A method for recommending music that includes identifying a granularity of a plurality of genres based on a request for music similarity, wherein the request identifies a user, training a genre classifier based on the granularity to obtain a trained genre classifier, calculating a first profile by the trained genre classifier, wherein the first profile that includes, for each of the plurality of genres, the likelihood that a music selection associated with a user is in the genre, calculating a second profile by the trained genre classifier, wherein the second profile that includes, for each of the plurality of genres, the likelihood that an unknown music selection is in the genre, obtaining a first similarity score between the first profile and a second profile, and recommending the unknown music selection to the user based on the first similarity score.